BRNO UNIVERSITY OF TECHNOLOGY
Faculty of Electrical Engineering and Communication
Department of Biomedical Engineering
Ivo Provazník, Ph.D.
WAVELET ANALYSIS FOR SIGNAL DETECTION -APPLICATIONS TO EXPERIMENTAL
CARDIOLOGY RESEARCH
VLNKOVÁ ANALÝZA PRO DETEKCI SIGNÁLŮ –APLIKACE V EXPERIMENTÁLNÍM
KARDIOLOGICKÉM VÝZKUMU
SHORT VERSION OF HABILITATION THESIS
Brno 2002
KLÍČOVÁ SLOVAvlnková analýza, spojitá vlnková transformace, detekce signálů, kardiologie
KEY WORDSwavelet analysis, continuous wavelet transform, signal detection, cardiology
MÍSTO ULOŽENÍ PRÁCEarchiv FEKT VUT v Brně
2002 I. Provazník
ISBN 80-214-2086-3
ISSN 1214-418X
3
CONTENTS
CONTENTS ______________________________________________________ 3
1 INTRODUCTION _______________________________________________ 5
1.1 Tools of time-frequency analysis ____________________________________________ 6
1.2 Time and frequency resolution ______________________________________________ 6
2 MULTIRESOLUTION AND WAVELET TRANSFORM_______________ 7
2.1 Multiresolution analysis ___________________________________________________ 8
2.2 Continuous wavelet transform ______________________________________________ 8
2.3 Wavelets for time-frequency localization______________________________________ 9
2.3.1 Real-valued wavelets_______________________________________________ 10
2.3.2 Complex-valued wavelets ___________________________________________ 10
3 WAVELET ANALYSIS FOR SIGNAL DETECTION ________________ 11
3.1 Detection of waves using CWT ____________________________________________ 11
3.1.1 Detection using envelope contour_____________________________________ 12
3.1.2 Basics of complex-valued CWT analysis _______________________________ 12
3.1.3 Detection of waves using modulus of complex-valued CWT ________________ 13
3.1.4 Detection of waves using phase of complex-valued CWT___________________ 14
3.2 Conclusions____________________________________________________________ 14
4 WAVELET ANALYSIS IN CARDIOLOGY RESEARCH _____________ 15
4.1 Detection of myocardial ischemia __________________________________________ 16
4.1.1 Electrophysiological manifestation of acute myocardial ischemia ___________ 16
4.1.2 t-test analysis of signal changes ______________________________________ 17
4.2 Hidden Markov model based detector of acute myocardial ischemia _______________ 19
5 WAVELET ANALYSIS IN EDUCATION __________________________ 19
6 CONCLUSIONS________________________________________________ 20
4
Ivo Provazník was born in 1968 in Brno. In 1991, he obtained
Ing. (MSc.) degree in radioelectronics and in 1996, Ph.D. degree
in electronics, both at the Brno University of Technology.
Since 1993, Ivo Provazník has been working as a lecturer at the
Department of Biomedical Engineering of Brno UT. He is
engaged in teaching programming, expert systems, and
advanced digital signal processing. His research interests
include time-frequency methods of biosignal analysis,
reconstruction and visualization of medical images.
5
1 INTRODUCTION
Today's signal processing tools have been developed through dozens of years to
cover all possible application areas - from data analysis to data compression. How-
ever, some specific but everyday technical problems cannot be effectively resolved
without techniques that are more complex. One of the newest additions has been
wavelets.
Although the subject area of wavelets has developed mostly over the last fifteen
years, it is connected to older ideas in many other fields including pure and applied
mathematics, physics, computer science, and engineering. Roots of "modern"
wavelet theory have been founded at several places in the late 1970's and in the
1980's. First, J. Morlet came up with an alternative for the short-time Fourier trans-
form. Morlet followed two aims: to gain high time-resolution for high frequency
transients and good frequency resolution for low frequency components. While
these two goals form a trade-off in traditional short-time Fourier transform, Morlet
decided to generate the transform functions in a different way: he took a windowed
cosine wave (using a smooth window) and compressed it in time to obtain higher
frequency function, or spread it out to obtain lower frequency function. In order to
examine time changes, these functions were shifted in time as well.
The first aim of this thesis is to describe tools of wavelet analysis with emphasis
to continuous-time wavelet transform. Thus, comprehensive overview on published
theory and practical results is presented. Second aim is to present details of features
of wavelet-based tools and present possibilities of their application. Third aim is to
show how the wavelet tools can be implemented in experimental cardiology re-
search.
In the thesis, the term wavelet analysis represents expansion of a discrete-time or
continuous-time signal on wavelet bases. Generally said, the expansion can be pro-
vided by any well-known signal processing tool such as Fourier transform.
However, wavelet analysis exploits a simple but genius idea - the signal is expanded
on a set of dilated or compressed functions
−
a
btψ .
The dilation a - the scale - is the key factor that allows to change both time and
frequency resolution when analyzing the signal. The signal can be analyzed to detect
short-time frequency-limited events, it can be more effectively compressed, or noise
present in the signal can be suppressed in time- and frequency-selective manner.
Wavelets have been extensively used in biomedical engineering since their first
formal formulation in the end of 1980's. First special journal issue1 on wavelets in
1 IEEE Engineering in Medicine and Biology Magazine 14(2), 1995 - Special Issue on Time-
Frequency and Wavelet Analysis.
6
biomedical engineering was released in March 1995. The guest editor M. Akay
chose outstanding papers on analysis of uterine EMG signals, EEG signals in epi-
lepsy, heart sound signals, ECG signals in ventricular fibrillation, fetal breathing
rate, respiratory-related evoked potentials, and others. Those papers showed wave-
lets in applications where classical analysis tools failed.
1.1 Tools of time-frequency analysis
Traditional frequency analysis by Fourier transform has many alternative methods.
The necessity of their application is provoked by non-stationary characteristics of
the signals being analyzed. Most published methods are short-time Fourier trans-
form, Wigner-Ville distribution, and wavelet transform.
Short-time Fourier transform is locally applied Fourier transform. The signal f(t)
is first multiplied by a shifted window function w(t-τ). Then, the conventional Fou-
rier transform is taken. The resulted transform is represented by a two-dimensional
function. A number of various window functions have been proposed to achieve
good-time frequency resolution. A good example is the use of Gaussian window
proposed by Gabor in 1946.
Wigner distribution or more often Wigner-Ville distribution is a well-known ex-
ample of expansion. Its intention is to get estimation of instantaneous power
spectrum. The attractive feature of Wigner-Ville distribution is the possible high
time-frequency resolution. For signals with a single time-frequency component the
Wigner-Ville distribution gives a clear and concentrated energy ridge in the time-
frequency plane. However, in the case of multicomponent signals, cross terms and
interferences appear.
Continuous wavelet transform (CWT) uses shifts and scales (dilation and con-
traction) of the prototype function ψ(t) instead of its shifts and modulations. CWT
gives conical pattern showing good frequency resolution for high scales corre-
sponding to low frequencies and poor frequency resolution for low scales
corresponding to high frequencies.
1.2 Time and frequency resolution
In any signal processing applications, time-frequency localization - i.e. localization
of a given basis function in time and frequency - is an important consideration. Sig-
nal domain methods require a high degree of localization in time while frequency
domain methods demand a high degree of localization in frequency. This results in
trade-off that can be optimized but not made ideal.
Definition of localization of a basis function is usually based on how it covers
certain area in time-frequency plane. The elementary area in the plane is called a
tile. Ideally, the tile is represented by a small rectangular window centered on the
place of interest in the time-frequency plane.
7
To center the tile in the plane, the transforms used for time-frequency representa-
tion use elementary operations such as shift in time, modulation in frequency, and
scaling. Obviously, shift in time by τ results in shifting of the tile by τ across the
time axis. Similarly, modulation by eiωst shifts the tile by ωs. All elementary opera-
tions conserve the area of the tile. In addition, note that the tile shape is never ideally
rectangular or never has infinitely narrow dimensions. Its real shape is determined
by basis functions used for expansion.
Consider a signal f(t) centered on t0 with its frequency spectrum F(ω) centered
on ω0. Let us define time resolution as time width ∆t of f(t) by its root mean square
spread and frequency resolution as frequency width ∆ω of F(ω) by square of the 2nd
moment of |F(ω)|2
( ) ( )∫∞
∞−
−=∆ dttfttE
t
22
0
2 1, (1.1)
( ) ( )∫∞
∞−
−=∆ ωωωω
πω
dFE
22
0
2
2
11, (1.2)
where E is energy of the signal. Resolutions in time and frequency are related in un-
certainty principle [6]. The principle sets a bound on the maximum theoretical joint
resolution in time and frequency represented by a product ∆t∆f. If f(t) decays faster
than t/1 as t→∞, then uncertainty principle asserts [19]
2
122≥∆∆
ωt. (1.3)
2 MULTIRESOLUTION AND WAVELET TRANSFORM
Wavelets arose in diverse scientific areas from mathematics to engineering and have
been described in various ways. They have been formalized in terms of multiresolu-
tion analysis and continuous wavelet transform, and further in discrete wavelet
transform and subband coding.
Multiresolution theory has been formulated in 1986 by S. Mallat [13] and
Y. Meyer. It provides a framework for understanding and description of wavelet
bases. The basic idea of the analysis is based on existence of a sequence of embed-
ded approximation spaces. If some six requirements that define the multiresolution
analysis are satisfied, then there exists an orthonormal basis on which a signal can
be expanded. Such expansion is non redundant, effective, and easily understandable.
In fact, its interpretation is not more difficult than interpretation of Fourier analysis.
Further, multiresolution analysis can be described in terms of subband coding using
multirate filter banks.
Continuous wavelet transform represents different approach but possesses further
desired features. Multiresolution theory - well described and ready to be imple-
8
mented - "only" allows (orthogonal) dyadic expansion that corresponds to signal
processing by an octave-band filter bank. Thus, frequency and time resolution is
preset and unchangeable. Opposite to it, continuous wavelet transform decomposes a
signal with arbitrary resolution in both time and frequency. Thus, CWT has received
significant attention in its ability to zoom in on singularities, which has made it an
attractive tool in the analysis of non-stationary and fractal signals. CWT can be dis-
cretized to obtain arbitrarily sampled time-frequency plane of wavelet coefficients.
2.1 Multiresolution analysis
Multiresolution theory based on multiresolution analysis and other mathematical
tools gives a foundation for exact description of expansion on orthogonal bases in a
Hilbert space L2(R). Let us use definition adopted by Daubechies in [6]: a multire-
solution analysis consists of a sequence of embedded closed subspaces
LL21012 −−
⊂⊂⊂⊂ VVVVV such that they posses the following properties: upward
completeness, downward completeness, scale invariance, shift invariance, and exis-
tence of an orthonormal basis ( ){ }Z∈− nnt |ϕ , for V0, where ϕ∈V0. The function
ϕ(t) is called the scaling function.
The scaling function composes an orthogonal basis for V0. Using the scale invari-
ance and the shift invariance, we can write orthonormal basis {ϕm,n}, n∈Z for the
space Vm [6]
( ) ( )nttmm
nm−=
−−
222/
, ϕϕ , for Z∈nm, . (2.1)
Daubechies also proves [6] existence of an orthonormal basis
( ) ( )nttmm
nm−=
−−
222/
, ψψ , for Z∈nm, , (2.2)
such that {ψm,n}, n∈Z, is an orthonormal basis for a space Wm, where Wm is the or-
thogonal complement of the space Vm in Vm-1. The function ψ(t) is called the
wavelet function.
2.2 Continuous wavelet transform
Let us consider redundant representation of continuous-time functions in terms of
two variables - scale and shift. The representation is called continuous wavelet trans-
form. Although the CWT is redundant, it is worth of use because of its interesting
features. Consider a family of functions obtained by shifting and scaling a wavelet
function ψ(t) such as
( )
−= ∗
a
bt
at
baψψ
1
,
, (2.3)
where a, b∈R, a>0, and a star denotes complex conjugate. The normalization en-
sures that ( ) ( )ttba
ψψ =,
. The wavelet function has to oscillate. This, together with
the decay property, has given ψ(t) the name wavelet or "small wave".
9
The continuous wavelet transform of a function f(t) is defined as (e.g. [6])
( ) ( ) ( )∫∞
∞−
= dttftbaCWTba ,
, ψ . (2.4)
The function f(t) can be recovered from its wavelet transform by the reconstruction
formula
( ) ( ) ( )∫ ∫∞
∞−
∞
∞−
= dbdatbaCWTaC
tfba,2
,11
ψ
ψ
. (2.5)
The CWT possesses properties similar to properties of other linear transforma-
tions: linearity, shift property, localization property, energy conservation, and
scaling property.
Because of high redundancy in continuous CWT(a,b), it is possible to discretize
the transform parameters. CWT often uses a hyperbolic grid in case of linear grid of
a. A special case of the hyperbolic grid is a dyadic grid when scales are powers of 2.
That is, the time-scale plane (a, b) is discretized as
m
aa0
= , m
abnb00
= , (2.6)
where m,n∈Z, a0>1, b0>0. In this manner, large basis functions (when is a0m
large) are shifted in large steps, while small basis functions are shifted in small
steps. In order for the sampling of the time-scale plane to be sufficiently fine, a0 has
to be chosen close to 1, and b0 close to 0. The discretized family of wavelets is now
( ) ( )002/
0, nbtaatmm
nm−=
−−
ψψ . (2.7)
2.3 Wavelets for time-frequency localization
Wavelets are basis functions used for expansion. They are characterized by a num-
ber of properties that determine their use in the frame of time-frequency localization.
Formally, a real-valued function ψ(t) is called a wavelet if it satisfies two constraints
defined by
( ) 0=∫∞
∞−
dttψ and ( ) 12
=∫∞
∞−
dttψ . (2.8)
The first part of Eq. 2.8 states that the wavelet oscillates, the second part says the
wavelet must be nonzero somewhere. The properties of wavelets may serve as a key
for selection of function for a specific application. Briefly, while analysis needs even
non-orthogonal wavelets, compression requires orthogonal and smooth wavelets.
Filtering may require symmetrical functions and rational coefficients of filters corre-
sponding to wavelets. The following properties are most discussed in literature:
orthogonality, compact (finite) support, rational coefficients of corresponding filters,
symmetry, smoothness, and analytic expression.
10
2.3.1 Real-valued wavelets
The most used and/or discussed real-valued wavelets are: Haar wavelet, family of
Daubechies wavelets, Morlet wavelet, Meyer wavelet, Mexican hat wavelet, family
of Coiflet wavelets, family of Symlet wavelets, and biorthogonal wavelets. Time and
frequency resolution of various wavelets differ. The ideal resolution value is repre-
sented by an equality curve ∆2
t∆2
ω=0.5. Results for all wavelets lay right and above
the equality curve. The closer to the equality curve, the better time resolution, fre-
quency resolution, or both resolutions are. The results for selected wavelets are
summarized in Tab. 2.1.
wavelet ∆2
t ∆2
ω ∆2
t ∆2
ω
Morlet 0.7071 0.7081 0.5007
Gaussian No.2 0.7637 0.6889 0.5261
Meyer 0.8418 0.9824 0.8271
Daubechies No.2 1.540 9.424 14.51
Haar 0.5775 130.6 75.44
Tab. 2.1 Time resolution, frequency resolution, and time-frequency resolution of selected
real-valued wavelets. Theoretical minimum of ∆2
t ∆2
ω is 0.5.
2.3.2 Complex-valued wavelets
The most used and/or discussed complex-valued wavelets are: Complex Gaussian
wavelets, Complex Daubechies wavelets, Complex Kingsbury wavelet, Complex
Morlet wavelets, Complex Frequency B-spline wavelets, Complex Shannon wave-
lets. Time and frequency resolution of various complex-valued wavelets differ too.
The results for selected wavelets are summarized in Tab. 2.2.
wavelet ∆2
t ∆2
ω ∆2
t ∆2
ω
cpx Morlet No.1-0.5 (real or imaginary part) 0.3533 1.416 0.5006
cpx Kingsbury (real part) 1.552 3.418 5.306
cpx Kingsbury (imaginary part) 1.565 3.681 5.765
cpx Daubechies No.6 (real part) 2.526 3.066 7.749
cpx Daubechies No.6 (imaginary part) 2.647 3.363 8.904
Tab. 2.2 Time resolution, frequency resolution, and time-frequency resolution of selected
complex-valued wavelets. Theoretical minimum of ∆2
t ∆2
ω is 0.5.
11
3 WAVELET ANALYSIS FOR SIGNAL DETECTION
Wavelet analysis of signals is modern approach to solution of many digital signal
processing problems nowadays. When proper wavelet tools are chosen and proper-
ties of the tools met conditions based on certain signal parameters, the wavelets can
powerfully serve to reach significantly "better" results compared to e.g. results of
Fourier transform methods.
Wavelet tools and all their variants result in 2-dimensional continuous or discre-
tized output of two parameters: scale and time shift. While time-shift corresponds to
time axis of the signal to be analyzed, scale corresponds, but is not equal, to fre-
quency. This makes interpretation of wavelet analysis less explicit than in the
Fourier analysis. In practice, wavelet analysis output is often called time-frequency
image or time-frequency spectrum. However, correct interpretation has to consider
frequency spectra of individual scaled wavelets.
"How-to" of wavelet signal analysis naturally comes out of the transform proper-
ties. Localization property, shift property, and scaling property compose the
fundamentals of all detection algorithms based on the wavelet transform. However,
when using wavelets, we are faced several problems that must be resolved before the
signal is analyzed.
First of all, we have to choose the following parameters of the wavelet analysis:
type of the wavelet transform, type of the mother wavelet, and scales. All the three
wavelet analysis parameters have several or even infinite number of options. The
parameters are closely related and must be considered at the same time.
The type of the wavelet transform is usually the first step. It leads to the general
decision whether to use orthogonal expansion or overcomplete expansion. Orthogo-
nal expansion usually leads to a bank of octave filters representing dyadic discrete-
time wavelet transform. Overcomplete expansion is usually represented by the con-
tinuous-time wavelet transform computed on a given grid to discretize the resulted
continuous-time function. The wavelet type should be set according to the signal
being analyzed. The general selection of wavelets is based on the shape of the
wavelet in time domain, its length (support), and smoothness. The last step is based
on how many details in what frequency range is needed. This is done by selection of
scales for what the analysis will be computed. The simplest case is the dyadic dis-
crete-time wavelet transform. Here, the scales are set to powers of two. The
selection is reduced to "how deep" the analysis will be done. In other words, we
have to select the number of the coarsest levels. Overcomplete CWT also allows us
to choose arbitrary scales according to actual signal properties, if known.
3.1 Detection of waves using CWT
Detection of waves and short-time events is an important part of signal analysis.
Thus, the signal can be examined to find differences from a reference signal, track
12
long-term trends, and multiple time overlapping and/or frequency overlapping
changes. Traditional time-domain and frequency-domain detection methods are
based on correlation and cross-correlation, coherence, cross-spectra, cepstra, and
many other signal processing tools. Time-frequency approach exploits expansion on
series to decompose the signal into multiple frequency bands. Further, time and fre-
quency resolution can be individually changed in the bands and thus the analysis
algorithm can be adapted to the signal being detected.
3.1.1 Detection using envelope contour
The user is faced various problems when CWT should be used to detect and bound
time events in wavelet signal analysis. The problems result from the nature of the
time-frequency image that is too complex for direct analysis. Originally, the output
of the CWT is two-dimensional and may be depicted by 3D plot, 2D shaded image,
2D contour image, and 1D plot of cross-sections through the output along time axis
or scale axis. A number of parameters can be observed in the images/plots: presence
and position of the peaks, slope of the peaks in various directions, etc.
The signal waves can be more precisely detected using envelope contours. First, a
contoured image of the output is taken. Further, square root of absolute value of the
output is taken to visualize more details by increasing the image dynamics. The im-
age is sliced at eight levels regularly spaced between zero and maximum of the
output function. The contour image CL(a,b) for set of M levels L is defined as
( )( )[ ]( )
+−∈
=otherwiseif
LLbaCWTabssqrtifbaC kk
L0
;,1,
εε
, (3.1)
for all k's, where k=1..M is a level number, Lk is a k-th level, ε is a small number.
The levels are defined as
( )[ ]k
M
baCWT
LBbAa
k*
,max, ∈∈
= , (3.2)
where A is a set (or an interval) of all considered scales, and B is a set of all signal
samples (or a time interval) of the analyzed signal. Second, only that part of the
contour L1, which is the closest to the highest frequency, is considered. Such a con-
tour is called an envelope contour EC and is defined as
( )( )
[ ]abECbaCAa L 0,,
1
min≠∈
= , (3.3)
for all b's. The envelope contour EC is represented by a 1D signal that can be ana-
lyzed by common detectors or recognizers.
3.1.2 Basics of complex-valued CWT analysis
Complex-valued wavelet transform plays a special role in signal analysis. Complex
nature of wavelets provides further improvement in signal detection compared to
13
real-valued wavelet analysis. This is possible by using so called dual-tree processing
[10] through cross-correlation with real and imaginary parts of wavelets. The re-
sulted complex-valued time-frequency image can be further analyzed by detection of
significant attributes in its modulus and phase. In this way, not only the waves can
be detected but also various shapes of the waves can be distinguished.
Let us present the complex-valued CWT abilities on an example. The tested sig-
nal has been artificially corrupted by a short-time event - a discontinuity in first
derivative at t=84 msec. Such discontinuity can hardly be seen in the time domain
without further processing.
Then, the signal has been transformed using the complex Morlet wavelet No. 1-
0.5. As the CWT promises to detect short-time events regardless their frequency
contents, we should obtain a significant pattern in a resulting time-frequency image.
Studying the modulus CWT output depicted in Fig. 3.1 (b), one can easily find a
narrow object (islet) located at scales 1/a=0.2-0.3 (mid-frequencies). Although the
islet is low in value, it is detectable with relatively good time resolution.
0 50 100 150 200-0.06
-0.04
-0.02
0
0.02
0.04
0.06
0.08
0.1
0.12
time [msec]
am
plit
ud
e [
-]
signal
discontinuity
0 50 100 150 2000
0.1
0.2
0.3
0.4
0.5
0.6
time [msec]
1/s
cale
[-]
modulus of CWT
detected
discontinuity
Fig. 3.1 (a) Signal with artificial discontinuity in first derivative at t=84 msec, (b)
modulus of CWT using complex Morlet wavelet No.1-0.5.
Complex-valued CWT analysis may be further improved by considering its phase.
The phase of the complex-valued CWT of signals has characteristic structure. It
contains phase discontinuities along time axis that reveals as vertical line objects.
The lines correspond to extrema and inflection points of waves. The discontinuity in
the signal is displayed as an additional line in the phase.
3.1.3 Detection of waves using modulus of complex-valued CWT
To study complex-valued CWT behavior on various signal shapes, let us consider a
first derivative of square modulus and phase. Thus, we may wish to detect extrema
and inflection points that clearly define particular waves of the signal. By approxi-
mating the derivative of square modulus of CWT, we can establish connections
between local extrema of |CWT(a,b)|2 and inflection points and local extrema of
a signal being analysed. |CWT(a,b)|2 is considered as a function of time shift b
14
here (with constant scale a). Local maxima (minima) of |CWT(a,b)|2 always rep-
resent inflection points (local minima or maxima) of the signal, respectively.
One can find that the maxima and minima partly overlay each other. To detect
extremal points or segments in the CWT, the following simple method can be used.
The method is based on detection of local maxima of the CWT modulus. A local
maximum of the CWT modulus located at time b2 at scale a is defined as
( ) ( ) ( )321
,,, baCWTbaCWTbaCWT >< , (3.4)
where b1<b2<b3, b1→b2 and b3→b2 (b1, b2, b3 are adjoining samples when the
signal being analysed is a discrete-time signal). If inequality in Eq. 3.4 is applied on
all a's, adjoining {a, b2} pairs compose separated "lines" called maximum curves.
The maximum curves coincide with the minor extrema of the function.
Further analysis of the modulus of the CWT is possible by detection of minima. A
local minimum of the CWT modulus located at time b2 at scale a is defined similar
to Eq. 3.4 with reversing the inequalities. If inequality equation is applied on all a's,
adjoining {a, b2} pairs compose separated "lines" called minimum curves.
3.1.4 Detection of waves using phase of complex-valued CWT
To study behavior of CWT of signals with local extrema or inflection points, we
consider the case of functions exhibiting local symmetry (or anti-symmetry) proper-
ties. CWT(a,b0) of locally symmetric f(t) is real and its phase is then 0 or π.
CWT(a,b0) of locally anti-symmetric f(t) is imaginary and its phase is then π/2 or
-π/2.
For analysis purposes, the phase can be thresholded to obtain a less complex
output image. Thresholded phase is defined as
( )( )[ ]
( )[ ]
+−∉
+−∈=
εε
εε
ththbaCWTfor
ththbaCWTforbaP
th,,arg0
,,arg1, , (3.5)
where th={-π/2, 0, π/2, π} is threshold, ε is a small real number. The phase
thresholding at {-π/2, 0, π/2, π} results in binary image with vertical "lines" lo-
cated at positions of phase steps.
3.2 Conclusions
Behavior of the complex-valued continuous wavelet transform (CWT) as response
to various signal types has been discussed. Experiments have shown that CWT may
serve as a detector of signal changes. We have proved that the complex-valued
wavelet transform can be used for detection of waves represented by local maxima
and minima of the signal. Further, it can be used for recognition of symmetric and
antisymmetric waves. As the first step of the complex-valued wavelet analysis in
detection, square modulus (or simply modulus) of CWT should be computed. Thus,
maxima, minima, or even inflection points are found as detectable maxima or min-
15
ima. Phase of CWT is used to distinct between extrema and inflection points in the
signal being analysed.
However, the examples discussed above represent ideal conditions that may be far
from reality. First, all signals are recorded with nonzero signal to noise ratio. The
noise may produce local extrema that disturb CWT and make the detection more
complicated. Further, no signal is exactly symmetric (antisymmetric). Even small
asymmetry degrades resulting peaks in CWT modulus.
Consider a repetitive signal of a series of waves. After a certain time delay, the
signal is corrupted by additive noise represented by short-time waves of lower value
than the signal amplitude. The noise partially overlaps the signal in time and fre-
quency domain. Comparing CWT of both original and corrupted signals, one can
find detectable differences in modulus as well as phase of CWT. The modulus re-
veals the differences as additional peaks in its image. Although the differences are
small in value, they change shape of original peaks in modulus image or they gener-
ate separated peaks. The phase responses even more sensitively regardless the noise
wave amplitude. Any new signal component is revealed as a new phase step along
time axis.
Analysis of signals using complex-valued continuous wavelet transform is the
first step to detect possible changes or alternans. In the second step, modulus and
phase must be thoroughly examined. However, the complex-valued time-frequency
image is too complex. Its further analysis may fail when using simple methods (po-
sition of main peaks, number of new peaks in modulus, number of new π-steps in
phase, etc).
4 WAVELET ANALYSIS IN CARDIOLOGY RESEARCH
Wavelet analysis has been linked to signal processing in early 1990's. A number of
applications to various fields have been described since that time. A typical exam-
ples are telecommunication, mechanical engineering, geology, climatology,
oceanology, astrophysics, computer science, and biomedical engineering.
Signal detection applications of wavelets in biomedical engineering include broad
class of tasks [4]. In cardiology research, wavelet analysis is exploited in electrocar-
diographic (ECG) signal compression, ventricular arrhythmia analysis, heart rate
variability analysis, cardiac pattern characterization, late potentials analysis, fetal
ECG extraction, heart sounds analysis, ventricular pressure variability, high-
resolution ECG analysis, detection of T-wave changes and ST-T complex changes,
detection of conduction block, and many others.
Wavelets are an efficient tool for analysis of short-time changes in signal mor-
phology. As pointed out by Unser and Aldroubi in [18], the preferred type of
wavelet transform for signal analysis is the redundant one that is the continuous
wavelet transform in opposition to the non-redundant type corresponding to the ex-
pansion on orthogonal bases (multiresolution analysis). The reason is that the CWT
16
allows decomposition on an arbitrary scale. Thus, frequency bands of interest can be
studied properly at chosen resolution.
In the following text, wavelet analysis is applied to cardiology. There, electrocar-
diographic signals are used to detect a pathological process in the heart (myocardial
ischemia). ECG signals are recorded from the same subject in two phases: control
(physiological signal) and after the event (artery occlusion causing acute ischemia).
Thus, the original signal and corrupted signal are taken and can be used to design an
effective detection algorithm. However, even the physiological signal is noisy it-
self. This is caused by dynamic nature of the signal source and the dynamic systems
between the source and recording electrodes. The analysed ECG signal is then com-
posed of repetitions that vary (beat-to-beat variations). Such variations may
negatively influence the detection and may increase a number of false positive
events (low specificity). The same negative effect may be caused by other unwanted
noise, e.g. powerline noise, moving artifacts, myopotentials, etc.
4.1 Detection of myocardial ischemia
In the Western world, sudden cardiac death remains a leading cause of death. In the
majority of the cases, sudden death is caused by lethal arrhythmia’s preceded by
acute myocardial ischemia. The study of ischemic heart disease can reveal mecha-
nisms of its genesis and its influence in the electrophysiology of the heart. Results of
the study could then contribute to a better understanding and both pre-infarction and
post-infarction treatment of the disease. Further, the results could help to develop a
new noninvasive method needed in cardiology diagnostics [7].
Here, wavelets are used to detect acute myocardial ischemia caused by occlusion
of a coronary artery. This application may help to understand fundamentals of elec-
trophysiological changes underlying myocardial ischemia.
4.1.1 Electrophysiological manifestation of acute myocardial ischemia
Coronary ischemia is characterized by interrelated metabolic ionic and neurohu-
moral events that alter membrane properties of cardiac cells, causing
electrophysiological changes. A deteriorated myocardial perfusion causes a potential
difference between ischemic and normal regions during the ST segment. While ST-
depression is considered the most common manifestation of exercise-induced car-
diac ischemia, ST-elevation may be related to severe posterior, subepicardial or
transmural ischemia, or with myocardial infarction. Sometimes, fixed ST elevation
or depression may occur. ST changes may therefore be ambiguous and are not al-
ways able to reflect changes in myocardial perfusion [8]. Concluding, cardiac
ischemia may result in changes in one or more leads of the electrocardiogram. The
electrocardiographic criterion for detecting ischemia is ST-segment displacement.
The value of this criterion for predicting coronary artery disease is limited and is re-
ported between 47 and 91% for sensitivity and between 69 and 97% for specificity.
17
Since ischemia causes conduction changes, irregular depolarization (activation) of
the myocardium may occur. This would be manifested as intra-QRS changes. There
is much evidence that ischemia changes in the heart muscle may cause alterations in
the QRS spectrum, as an expression of the fragmentation of ventricular depolariza-
tion. Abboud [1] detected high-frequency changes in signal-averaged QRS
complexes of dogs and human patients caused by ischemia. Further, focal reduction
of high-frequency components of the QRS complex under myocardial ischemia in-
duced by percutaneous transluminal coronary angioplasty have been showed [8].
Therefore, a technique similar to the spectrotemporal analysis of late potentials [12]
might prove useful in early detection of ischemic changes. This idea is further pro-
moted by Petterson [15], where ST-segment analysis criteria are combined with
root-mean-square values of QRS high-frequency components criteria.
4.1.2 t-test analysis of signal changes
The main issue in analysis of QRS changes is to localize the abnormal time-scale
components contained in the ECG signal to identify a given cardiac disease. A
method compares the representations of the ECGs of the studied (ischemic) sample
to a reference (control) sample. The significant abnormality mapping is assessed by
comparing the mean value of each of the wavelet transforms of the two studied
populations by means of an two-way two-tailed t-test to test for the null hypothesis
that means are equal. If the null hypothesis is rejected, the statistically significant
appearance of QRS changes in time-frequency spectra is confirmed.
Data were collected under the following protocol. The signals of one-minute
length were recorded in 15 time instances from three orthogonal leads (X, Y, Z).
Thus, a record from control period (0 min) and six records from ischemic period (1,
3, 5, 10, 15, 20 min) were taken. A typical signal recorded from X-lead is depicted
in Fig. 4.1. One can see several time-domain changes during all periods. Comparing
to control period, ST-segment elevates in ischemic period. Further, T-wave increase
in amplitude and shifts to QRS-complex during ischemic period. Other minor
changes are visible too.
0 min 1 min 3 min 5 min 10 min
Fig. 4.1 Typical recordings from myocardial ischemia experiment. Legend: 0 min -
baseline recording, 1 min to 10 min - acute ischemia.
The data were reduced to set null hypothesis for statistical t-test. Two recordings
were chosen to generate studied sample and reference sample. Studied sample were
18
composed of recordings after 3 minutes of ischemia, the reference sample were
composed of recordings from control period. L consequent heart cycles were chosen
in each recording of all n experiments. Heart cycles are physiologically of (slightly)
different length. Therefore, M samples centered on a fiducial point FP were chosen.
FP was set as an arithmetic center between QRS-onset and QRS-offset of each heart
cycle. Thus, n*L*2 recordings of heart cycles were taken for the test. Overall num-
ber of signal samples for statistical analysis was n*L*2*M. The results in this
chapter are shown for n=15, L=10, M=250 (i.e. 500 msec for sampling rate of
500 Hz).
Time-frequency manifestation of myocardial ischemia should be graphically pre-
sented before some preprocessing and statistical analysis of wavelet data will be
done. Such presentation should prove that statistical analysis will likely reject null
hypothesis and that time-frequency image bear significant information on electro-
physiological changes due to acute myocardial ischemia.
0 min 1 min 3 min 5 min 10 min
Fig. 4.2 Continuous wavelet transform of recordings from Fig. 4.1 using Kingsbury
wavelet. 0 min - baseline recording, 1 min to 10 min - acute ischemia.
Fig. 4.2 shows CWT of all recordings from Fig. 4.1. Time-frequency images re-
veal some changes within QRS complexes during ischemia (center part of particular
pictures). Further, energy dissipation within QRS complex and energy accumulation
in T-wave is obvious.
It should be pointed out that QRS complex shape and duration may physiologi-
cally vary. These variations are small but may significantly influence statistical test.
Therefore, some filtering method should be applied. An efficient algorithm uses me-
dian filtering which computes median QRS complex from a set of L consequent
QRS complexes. Each of n*2 median QRS complexes were transformed using
CWT. Thus, n*2 matrices of M x S wavelet coefficients were computed. S is a
number of scales of CWT. These matrices provided input data for two-tailed t-test.
The CWT analysis discussed above using statistical results is likely a good ische-
mic marker. Sufficient efficiency was achieved for recordings from X-lead and in
CWT modulus (p<0.001 for larger areas within QRS complexes in time-frequency
images). The results can be used to develop an automatic on-line detector of acute
myocardial ischemia. The detector would be based on comparison of a current time-
frequency image to time-frequency image computed from a signal recorded at the
beginning of diagnostic procedure.
19
4.2 Hidden Markov model based detector of acute myocardial ischemia
Another detector uses two discrete density hidden Markov models. First model was
trained on a set of CWT images of control QRS complexes. Second model was
trained on a set of CWT images of QRS complexes from ischemic period. Both
models were build as ten-state left-right structures.
The left-right property of the model was chosen to follow time-nature of an ECG
signal where samples of the signal follow each other in one direction. The number of
states was experimentally chosen as usual in Markov model applications.
Individual states were represented by index of attribute vectors. Attribute vectors
are simply generated by taking vectors of M wavelet coefficients for each time in-
stant. Indexes of attribute vectors were set according to a codebook, which consisted
of minimized number of representative attribute vectors generated during training
phase.
The detector based on the above hidden Markov model was tested on a set of sig-
nals from 11 experiments where three vectorcardiograms were recorded before LAD
occlusion (control period) and after 3 minutes of acute ischemia (ischemic period).
One QRS complex of each cardiogram was included into analysis. The size of the
codebook was 50 attribute vectors. Number of states of the model was experimen-
tally set up to ten. Six models representing non-ischemic and ischemic state in
vectorcardiograms from lead X, Y, and Z were built up. Each signal was trans-
formed by CWT using Morlet wavelet and processed by two corresponding models.
The results for lead X are shown in the Tab. 4.1. The first row in the Tab. 4.1 repre-
sents that 81.8% control signals recorded from X-lead were recognized as control
signals. The second row represents that 90.9% ischemic signals recorded from X-
lead were recognized as ischemic signals.
lead tested signal used model score [%]
control control 81.8X
ischemic ischemic 90.9
Tab. 4.1 Results of myocardial ischemia detection using ten-state hidden Markov model.
5 WAVELET ANALYSIS IN EDUCATION
Wavelet analysis has undergone significant growth in the past few years, with many
successes in the efficient analysis, processing, and compression of signals and im-
ages. It is based on the idea of frequency-scale decomposition of signals and images,
which offers many advantages over the traditional frequency decompositions. Un-
fortunately, the new technique requires broad mathematical framework and adequate
knowledge of digital signal processing techniques. Thus, a comprehensive course
20
that includes fundamentals of multiresolution analysis, continuous and discrete-time
wavelet transform with appropriate examples is needed for potential users.
Inclusion of the topic into the university teaching system is possible when suffi-
cient fundamentals of signal processing and Fourier theory are given. A good
example is a set of two consequent one-semester courses "Signals and Systems" and
"Digital Signal Processing". Then, a new course of a given topic can follow either in
graduate or post-graduate study. The objectives of the course are as follows: i) to
give an introduction to wavelet analysis in one and two dimensions, and ii) show
that it may be practically used in applications of filtering and compression.
Wavelet analysis has already been partially included into teaching process at De-
partment of Biomedical Engineering, BUT. First, multirate signal processing, half-
band filters, filter banks, and discrete-time wavelet transform are covered in an op-
tional course "New Methods of Signal Processing" supervised and lectured by Dr.
Jiří Kozumplík, computer laboratory exercises led by Dr. Ivo Provazník. Second, a
number of semestral projects on wavelet analysis have been supervised by Dr. Ivo
Provazník. Although there is limited time for solving such projects during regular
semester, students are able to learn necessary wavelet basics. Very good knowledge
of digital signal processing including Fourier theory is necessary. Most of the se-
mestral projects have been continued and completed as Master's theses in 1996-
2001.
Third, a chapter on wavelet transform has been included into second edition of the
book [9] by J. Jan: Digital Signal Filtering, Analysis and Restoration and the basic
principles of wavelet transform are presently taught in the graduate course "Digital
Signal Processing".
6 CONCLUSIONS
Wavelets have generated an enormous interest in both theoretical and applied sci-
ence areas, especially over the past ten years. New advancements in the science are
occurring at such a rate that even the meaning of the term "wavelet analysis" con-
tinuously keeps changing to incorporate all new ideas. In the thesis, wavelet analysis
is discussed in terms of multiresolution analysis as a framework for orthogonal ex-
pansion of functions and series.
The text summarizes fundamentals of wavelet theory: basic mathematical back-
ground is given for basic understanding of time-frequency representation, common
time-frequency analysis tools are presented such as wavelet transform and short-
time Fourier transform, and time-frequency resolution is defined in frame of basis
function of time-frequency decomposition. Further, three roots of time-frequency
analysis are discussed: multiresolution analysis, subband coding and filter banks,
and continuous wavelet transform (CWT). Concerning the wavelet transform, a
comprehensive overview of important wavelets. Both real-valued and complex-
valued wavelets are discussed.
21
The thesis presents a novel contribution to wavelet-based detection of signals us-
ing complex-valued wavelets. The first technique exploits a localization property. A
simple technique generates an envelope contour of narrow signal elements and thus
the CWT analysis result into one-dimensional domain. The envelope contour is suit-
able to be used by common detectors of extrema. The second technique is based on
analysis of CWT modulus that provides information on local extrema and inflection
points of the analysed signal. The technique uses local maxima of the CWT modulus
and tracks CWT modulus ridges. Thus, so-called maximum curves are generated.
Their position and length may serve for further detection. Another technique is
based on analysis of CWT phase that provides further information on inflection
points of the analysed signal. Using a thresholding algorithm, curves of break phases
are generated. These curves may complete maximum curves for more effective de-
tection.
The study is completed by examples of two applications of CWT in cardiology
research. The first example shows how modulus of complex-valued CWT can be
used in detection of myocardial ischemia. A novel algorithm employing a hidden
Markov model working on time-scale image of recorded ECG signals is described.
Further, the model finds electrophysiological intra-QRS changes, which is a novel
technique compared to classical ST-segment analysis. The second example shows
how modulus of complex-valued CWT can be used in detection of electrophysi-
ological changes in the heart caused by neurological drugs. A simple algorithm
tracking two highest peaks in the time-scale image is discussed.
22
List of principal references
[1] Abboud S. High Frequency ECG - A New Method to Examine Depolarization Changes
Mediated by Transient Myocardial Ischemia. Proc Computers in Cardiology, 105-108, 1989.
[2] Akay M. Time-frequency and Wavelets in Biomedical Engineering. IEEE Press, 1997.
[3] Aldroubi A, Unser M. Wavelets in Medicine and Biology. CRC Press, New York, 1996.
[4] Bronzino J D. The Biomedical Engineering Handbook. CRC Press, 1997.
[5] Cohen A. Hidden Markov Models in Biomedical Signal Processing. Proc Int Conf IEEE
EMBS Vol. 20, No.3 , pp.1145-1150, 1998.
[6] Daubechies I. Ten Lectures on Wavelets. SIAM, Philadelphia, 1992.
[7] de Luna A B, Stern S. Future of Noninvasive Electrocardiology. In: Zareba W, Maison-
Blanche P, Locati E H. Noninvasive Electrocardiology in Clinical Practice, Futura
Publishing Co., pp.475-478, 2001.
[8] Gramatikov B, Brinker J, Yi-chun S, Thakor N V. Wavelet Analysis and Time-Frequency
Distributions of the Body Surface ECG Before and After Angioplasty. Computer Methods
and Programs in Biomedicine 62:87–98, 2000.
[9] Jan J. Digital Signal Filtering, Analysis and Restoration. IEE London, United Kingdom,
2000.
[10] Kingsbury N G. Image Processing with Complex Wavelets. Philosophy Transactions Royal
Society London Annals, to be published in 2001.
[11] Kleber A G, Fleischhauer J, Cascio WE. Ischemia-Induced Propagation Failure in the Heart.
In: Zipes D P, Jalife J. Cardiac Electrophysiology. W. B. Saunders Co., 174-181, 1995.
[12] Lander P, Gomis P, et al. Analysis of High-Resolution ECG Changes During Percutaneous
Transluminal Coronary Angioplasty. Journal of Electrocardiology 28(Suppl):39-40, 1995.
[13] Mallat S G. A Theory of Multiresolution Signal Decomposition: the Wavelet Representation.
IEEE Transactions on Pattern Analysis and Machine Intelligence 11(7):674-693, 1989.
[14] Okajima M, Kawaguchi T, Suzuki S. Detection of Higher Frequency Components at Mid-
QRS Stage of Electrocardiogram. IEEE Computer Society Press, pp.343–346, 1990.
[15] Pettersson J, Pahlm O, Carro E, Edenbrandt L, Ringborn M, Sörnmo L, Warren S G, Wagner
G S. Changes in High-Frequency QRS Components Are More Sensitive Than ST-segment
Deviation for Detecting Acute Coronary Artery Occlusion. Journal of the American College
of Cardiology 36(6):1827-1834, 2000.
[16] Sendhadji L, Thoraval L, Carrault G. Continuous Wavelet Transform: ECG Recognition
Based on Phase and Modulus Representation and Hidden Markov Models. In: Aldroubi A,
Unser M. (Ed.) Wavelets in Medicine and Biology. Crc Press, New York, 1996.
[17] Strang G, Nguyen T. Wavelets and Filter Banks. Wellesley-Cambridge Press, Wellesley,
1996.
[18] Unser M, Aldroubi A. A Review of Wavelets in Biomedical Applications. Proceedings of
the IEEE 84(4):626-638, 1999.
[19] Vetterli M, Kovačević J. Wavelets and Subband Coding. Prentice-Hall PTR, New Jersey,
1995.
[20] Yakubo S, Ozawa Y, et al. Intra-QRS High-Frequency ECG Changes with Ischemia. Is It
Possible to Evaluate These Changes Using the Signal-Averaged Holter ECG in Dogs?
Journal of Electrocardiology 28(Suppl): 234-8, 1995.
23
Souhrn
Algoritmy zpracování signálů byly vyvíjeny po několik posledních desetiletí pro řa-
du aplikačních oblastí - od předzpracování přes analýzu signálů až po kompresi dat.
Některé specifické, ale časté problémy nemohou být efektivně řešeny bez nových
sofistikovanějších technik. Jednou z nejnovějších významných metod, které přispí-
vají do této oblasti, jsou vlnky (orig. wavelets).
Prvním cílem habilitační práce je podat náhled na problematiku vlnkové analýzy
s důrazem na spojitou vlnkovou transformaci (CWT). V práci je prezentován ucele-
ný popis publikované teorie s uvedením praktických výsledků. Druhým cílem práce
je detailně popsat vlastnosti nástrojů vlnkové analýzy a naznačit možnosti jejich
praktického využití. Třetím cílem je uvedení praktických implementací metod vln-
kové analýzy v experimentálním kardiologickém výzkumu.
Pojem vlnková analýza v práci reprezentuje rozvoj signálů na bázi vlnek. Rozvoj
může být proveden např. Fourierovou transformací. Vlnková analýza ale využívá
jednoduché, avšak účinné myšlenky - transformačním jádrem je dilatovaná mateřská
vlnková funkce. Faktorem dilatace je tzv. škála, která určuje časové i frekvenční
rozlišení při analýze signálu. Takto mohou být v signálech detekovány krátkodobé
úzkopásmové jevy, signály mohou být efektivně komprimovány, nestacionární šum
může být potlačen s časovou i frekvenční selekcí, atd.
Teze obsahují nové příspěvky k vlnkové detekci signálů s použitím komplexních
vlnkových funkcí. První metoda využívá lokalizační vlastnosti CWT. Jednoduchý
algoritmus generuje tzv. konturovou obálku elementů analyzovaného signálu a vý-
sledek časově-frekvenční analýzy tak převádí do jednorozměrné oblasti. Konturová
obálka je snadno zpracovatelná běžnými detektory extrémů. Druhá metoda je zalo-
žena na analýze modulu časově-frekvenčního spektra. Modul CWT poskytuje
spolehlivou informaci o lokálních extrémech a inflexních bodech analyzovaného
signálu. Popisovaná metoda využívá lokálních maxim CWT. Takto jsou generovány
tzv. "maximální křivky", jejichž pozice a délka slouží pro detekci. Poslední popiso-
vaná metoda je založena na analýze fáze CWT, která poskytuje další informace o
vlastnostech signálu v inflexních bodech. Prahováním fáze CWT signálu jsou gene-
rovány tzv. křivky fázových zlomů, které přispívají ke zvýšení efektivnosti detekce.
Habilitační práce je doplněna podrobně popsanými příklady dvou aplikací spojité
vlnkové transformace v kardiologickém výzkumu. První příklad ukazuje, jak je mo-
dul komplexní CWT použit k detekci myokardiální ischemie. Nový algoritmus
využívá skrytý Markovův model pracující s časově-frekvenčním obrazem zazname-
naného EKG signálu. Model detekuje elektrofyziologické změny projevující se
uvnitř QRS komplexu, což představuje nový přístup v porovnání s tradiční analýzou
ST segmentu. Druhý příklad ukazuje použití analýzy modulu komplexní CWT pro
detekci elektrofyziologických změn v srdci způsobených vlivem neurologických lé-
ků. Zde je popsán jednoduchý algoritmus založený na sledování dvou nejvyšších
vrcholů časově-frekvenčního obrazu.